LDS Bank Conflict Expert
Use the production predictor as the single source of conflict semantics. Do not derive a second bank formula in this skill or in downstream analysis.
Workflow
- Collect the exact target, opcode, wave size, and one access record per logical LDS access. Give every access a stable ID, lane, byte address, and width.
- Read references/index.md for input routing and supported outputs. Read references/model-contract.md when constructing access records or interpreting conflict groups.
- Read the matching architecture reference linked from references/index.md. Reject a target that has no registered profile.
- From this skill directory, run
python scripts/predict.py request.jsonusing an environment whererockeis importable. See CLI invocation for standard input and installed-test usage. Treat diagnostics as part of the result, not as optional console commentary. - Analyze classifications, group membership, multiplicity, and diagnostics from that semantic document without recomputing them.
- Read references/comparison.md before comparing profiles and references/validation-boundaries.md before making architecture or validation claims.
Required boundaries
- Select exact profile identities. Never substitute
gfx90afor an unknown target. - Keep prediction CPU-only and independent of profiler output or private data.
- Distinguish broadcast, distinct-address conflict, inactive access, and invalid or unsupported input.
- Report empirical collision semantics only within the documented opcode, wave, and address scope. Do not infer physical organization from the model.
- Use public target names and qualitative semantics. Do not add internal links, paths, raw measurements, job metadata, or software performance figures.